RT-Lynx: Putting the GEMM Sparsity In a Right Way for Diffusion Models

Diffusion Transformers (DiT) achieve strong performance in image generation but incur substantial inference costs. While prior work has reduced this cost via quantization and distillation, semi-structured sparsity, which can nearly halve FLOPs, remains underexplored. A key reason is that most existing approaches focus on weight sparsification, and pruning 50% of the weights can remove critical model capacity and degrade generation quality. Our study, however, shows that DiT activations are intrinsically sparse and significantly more robust to N:M semi-structured sparsification than weights. Motivated by this observation, we advocate a paradigm shift from weight sparsification to activation sparsification. We propose RT-Lynx, which applies N:M sparsification to activations and incorporates error-compensation techniques to mitigate accuracy loss. We further implement highly optimized CUDA kernels tailored to this setting, achieving up to a 1.55x speedup on average in linear layers. Extensive experiments across multiple diffusion models demonstrate that our method preserves the generation quality of the original models while substantially accelerating inference.

Rethinking Cross-Layer Information Routing in Diffusion Transformers

Diffusion Transformers (DiTs) have become a de facto backbone of modern visual generation, and nearly every major axis of their design -- tokenization, attention, conditioning, objectives, and latent autoencoders -- has been extensively revisited. The residual stream that governs how information accumulates across layers, however, has been directly inherited from the original Transformer. In this paper, we present a systematic empirical analysis of cross-layer information flow in DiTs, jointly along depth and denoising timestep, and identify three concrete symptoms of traditional residual addition, namely monotonic forward magnitude inflation, sharp backward gradient decay, and pronounced block-wise redundancy. Motivated by this diagnosis, we propose Diffusion-Adaptive Routing (\textsc{DAR}), a drop-in residual replacement that performs \emph{learnable, timestep-adaptive, and non-incremental} aggregation over the history of sublayer outputs. Moreover, the proposed \textsc{DAR} is compatible with many modern Transformer enhancement methods, such as REPA. On ImageNet $256\times256$, \textsc{DAR} improves SiT-XL/2 by $2.11$ FID ($7.56$ vs.\ $9.67$) and matches the baseline's converged quality with $8.75\times$ fewer training iterations. Stacked on top of REPA, it yields a $2\times$ training acceleration in the early stage, suggesting cross-layer information routing as an underexplored design axis in diffusion modeling, one that operates orthogonally to existing representation-alignment objectives. Beyond pretraining, \textsc{DAR} can also be applied during the fine-tuning stage of large-scale T2I models and preserves high-frequency details during Distribution Matching Distillation.

Shiva-DiT: Residual-Based Differentiable Top-$k$ Selection for Efficient Diffusion Transformers

Diffusion Transformers (DiTs) incur prohibitive computational costs due to the quadratic scaling of self-attention. Existing pruning methods fail to simultaneously satisfy differentiability, efficiency, and the strict static budgets required for hardware overhead. To address this, we propose Shiva-DiT, which effectively reconciles these conflicting requirements via Residual-Based Differentiable Top-$k$ Selection. By leveraging a residual-aware straight-through estimator, our method enforces deterministic token counts for static compilation while preserving end-to-end learnability through residual gradient estimation. Furthermore, we introduce a Context-Aware Router and Adaptive Ratio Policy to autonomously learn an adaptive pruning schedule. Experiments on mainstream models, including SD3.5, demonstrate that Shiva-DiT establishes a new Pareto frontier, achieving a 1.54$\times$ wall-clock speedup with superior fidelity compared to existing baselines, effectively eliminating ragged tensor overheads.

RazorAttention: Efficient KV Cache Compression Through Retrieval Heads

The memory and computational demands of Key-Value (KV) cache present significant challenges for deploying long-context language models. Previous approaches attempt to mitigate this issue by selectively dropping tokens, which irreversibly erases critical information that might be needed for future queries. In this paper, we propose a novel compression technique for KV cache that preserves all token information. Our investigation reveals that: i) Most attention heads primarily focus on the local context; ii) Only a few heads, denoted as retrieval heads, can essentially pay attention to all input tokens. These key observations motivate us to use separate caching strategy for attention heads. Therefore, we propose RazorAttention, a training-free KV cache compression algorithm, which maintains a full cache for these crucial retrieval heads and discards the remote tokens in non-retrieval heads. Furthermore, we introduce a novel mechanism involving a "compensation token" to further recover the information in the dropped tokens. Extensive evaluations across a diverse set of large language models (LLMs) demonstrate that RazorAttention achieves a reduction in KV cache size by over 70% without noticeable impacts on performance. Additionally, RazorAttention is compatible with FlashAttention, rendering it an efficient and plug-and-play solution that enhances LLM inference efficiency without overhead or retraining of the original model.

EasyQuant: An Efficient Data-free Quantization Algorithm for LLMs

Large language models (LLMs) have proven to be very superior to conventional methods in various tasks. However, their expensive computations and high memory requirements are prohibitive for deployment. Model quantization is an effective method for reducing this overhead. The problem is that in most previous works, the quantized model was calibrated using few samples from the training data, which might affect the generalization of the quantized LLMs to unknown cases and tasks. Hence in this work, we explore an important question: Can we design a data-independent quantization method for LLMs to guarantee its generalization performance? In this work, we propose EasyQuant, a training-free and data-independent weight-only quantization algorithm for LLMs. Our observation indicates that two factors: outliers in the weight and quantization ranges, are essential for reducing the quantization error. Therefore, in EasyQuant, we leave the outliers (less than 1%) unchanged and optimize the quantization range to reduce the reconstruction error. With these methods, we surprisingly find that EasyQuant achieves comparable performance to the original model. Since EasyQuant does not depend on any training data, the generalization performance of quantized LLMs is safely guaranteed. Moreover, EasyQuant can be implemented in parallel so that the quantized model could be attained in a few minutes even for LLMs over 100B. To our best knowledge, we are the first work that achieves almost lossless quantization performance for LLMs under a data-independent setting and our algorithm runs over 10 times faster than the data-dependent methods.

MKQ-BERT: Quantized BERT with 4-bits Weights and Activations

Recently, pre-trained Transformer based language models, such as BERT, have shown great superiority over the traditional methods in many Natural Language Processing (NLP) tasks. However, the computational cost for deploying these models is prohibitive on resource-restricted devices. One method to alleviate this computation overhead is to quantize the original model into fewer bits representation, and previous work has proved that we can at most quantize both weights and activations of BERT into 8-bits, without degrading its performance. In this work, we propose MKQ-BERT, which further improves the compression level and uses 4-bits for quantization. In MKQ-BERT, we propose a novel way for computing the gradient of the quantization scale, combined with an advanced distillation strategy. On the one hand, we prove that MKQ-BERT outperforms the existing BERT quantization methods for achieving a higher accuracy under the same compression level. On the other hand, we are the first work that successfully deploys the 4-bits BERT and achieves an end-to-end speedup for inference. Our results suggest that we could achieve 5.3x of bits reduction without degrading the model accuracy, and the inference speed of one int4 layer is 15x faster than a float32 layer in Transformer based model.

PASTO: Strategic Parameter Optimization in Recommendation Systems -- Probabilistic is Better than Deterministic

Real-world recommendation systems often consist of two phases. In the first phase, multiple predictive models produce the probability of different immediate user actions. In the second phase, these predictions are aggregated according to a set of 'strategic parameters' to meet a diverse set of business goals, such as longer user engagement, higher revenue potential, or more community/network interactions. In addition to building accurate predictive models, it is also crucial to optimize this set of 'strategic parameters' so that primary goals are optimized while secondary guardrails are not hurt. In this setting with multiple and constrained goals, this paper discovers that a probabilistic strategic parameter regime can achieve better value compared to the standard regime of finding a single deterministic parameter. The new probabilistic regime is to learn the best distribution over strategic parameter choices and sample one strategic parameter from the distribution when each user visits the platform. To pursue the optimal probabilistic solution, we formulate the problem into a stochastic compositional optimization problem, in which the unbiased stochastic gradient is unavailable. Our approach is applied in a popular social network platform with hundreds of millions of daily users and achieves +0.22% lift of user engagement in a recommendation task and +1.7% lift in revenue in an advertising optimization scenario comparing to using the best deterministic parameter strategy.

On the geometry of generalization and memorization in deep neural networks

Understanding how large neural networks avoid memorizing training data is key to explaining their high generalization performance. To examine the structure of when and where memorization occurs in a deep network, we use a recently developed replica-based mean field theoretic geometric analysis method. We find that all layers preferentially learn from examples which share features, and link this behavior to generalization performance. Memorization predominately occurs in the deeper layers, due to decreasing object manifolds' radius and dimension, whereas early layers are minimally affected. This predicts that generalization can be restored by reverting the final few layer weights to earlier epochs before significant memorization occurred, which is confirmed by the experiments. Additionally, by studying generalization under different model sizes, we reveal the connection between the double descent phenomenon and the underlying model geometry. Finally, analytical analysis shows that networks avoid memorization early in training because close to initialization, the gradient contribution from permuted examples are small. These findings provide quantitative evidence for the structure of memorization across layers of a deep neural network, the drivers for such structure, and its connection to manifold geometric properties.

Syntactic Perturbations Reveal Representational Correlates of Hierarchical Phrase Structure in Pretrained Language Models

While vector-based language representations from pretrained language models have set a new standard for many NLP tasks, there is not yet a complete accounting of their inner workings. In particular, it is not entirely clear what aspects of sentence-level syntax are captured by these representations, nor how (if at all) they are built along the stacked layers of the network. In this paper, we aim to address such questions with a general class of interventional, input perturbation-based analyses of representations from pretrained language models. Importing from computational and cognitive neuroscience the notion of representational invariance, we perform a series of probes designed to test the sensitivity of these representations to several kinds of structure in sentences. Each probe involves swapping words in a sentence and comparing the representations from perturbed sentences against the original. We experiment with three different perturbations: (1) random permutations of n-grams of varying width, to test the scale at which a representation is sensitive to word position; (2) swapping of two spans which do or do not form a syntactic phrase, to test sensitivity to global phrase structure; and (3) swapping of two adjacent words which do or do not break apart a syntactic phrase, to test sensitivity to local phrase structure. Results from these probes collectively suggest that Transformers build sensitivity to larger parts of the sentence along their layers, and that hierarchical phrase structure plays a role in this process. More broadly, our results also indicate that structured input perturbations widens the scope of analyses that can be performed on often-opaque deep learning systems, and can serve as a complement to existing tools (such as supervised linear probes) for interpreting complex black-box models.

1-bit LAMB: Communication Efficient Large-Scale Large-Batch Training with LAMB's Convergence Speed

To train large models (like BERT and GPT-3) with hundreds or even thousands of GPUs, the communication has become a major bottleneck, especially on commodity systems with limited-bandwidth TCP interconnects network. On one side large-batch optimization such as LAMB algorithm was proposed to reduce the number of communications. On the other side, communication compression algorithms such as 1-bit SGD and 1-bit Adam help to reduce the volume of each communication. However, we find that simply using one of the techniques is not sufficient to solve the communication challenge, especially on low-bandwidth Ethernet networks. Motivated by this we aim to combine the power of large-batch optimization and communication compression, but we find that existing compression strategies cannot be directly applied to LAMB due to its unique adaptive layerwise learning rates. To this end, we design a new communication-efficient algorithm, 1-bit LAMB, which introduces a novel way to support adaptive layerwise learning rates even when communication is compressed. In addition, we introduce a new system implementation for compressed communication using the NCCL backend of PyTorch distributed, which improves both usability and performance compared to existing MPI-based implementation. For BERT-Large pre-training task with batch sizes from 8K to 64K, our evaluations on up to 256 GPUs demonstrate that 1-bit LAMB with NCCL-based backend is able to achieve up to 4.6x communication volume reduction, up to 2.8x end-to-end speedup (in terms of number of training samples per second), and the same convergence speed (in terms of number of pre-training samples to reach the same accuracy on fine-tuning tasks) compared to uncompressed LAMB.